Introductory Dynamic Macroeconomics

Transcription

1 ii Introductory Dynamic Macroeconomics Ragnar Nymoen 10 January 2005

2 iv CONTENTS APhillipscurvemodel An error correction model that integrates the main-course.. 64 A Variables and relationships in logs 67 Contents Preface v 1 Dynamic models in macroeconomics Introduction Static and dynamic models, an example Dynamicmultipliers Dynamic effects of increased income on consumption GeneralnotationoftheADLmodel Atypologyoflinearmodels Extensionsandexamples Extensions Afewmoreexamples Multipliers in the text books to this course Theerrorcorrectionmodel Staticmodelsreconsidered Solution and simulation of dynamic models SolutionofADLequations Simulationofdynamicmodels Dynamicsystems Wage-price dynamics Introduction The main-course model of open economy inflation A framework for long-run wage and price setting Dynamicadjustment A simulation model of the main-course Themain-courseandthepricelevel The main-course model and the Scandinavian model of inflation The main-course model and the Battle of the Mark-ups Roleofexchangerateregime Themain-courseandthePhillipscurve Summing up the main-course model Norwegian evidence on the main-course model iii

3 vi PREFACE Preface This note is written for the course ECON 3410 /4410 International macroeconomics and finance. It reviews the key concepts and models needed to start addressing economic dynamics in a systematic way. The level of mathematics used does not go beyond simple algebra. References to the textbooks by Burda and Wyplosz B&W hereafter) and Rødseth s Open Economy Macroeconomics (OEM hereafter) are integrated several places in the text. v

4 2 CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS 75 The Norwegian current account Chapter 1 Billion kroner Dynamic models in macroeconomics 1.1 Introduction In many areas of economics, time plays an important role: firms and households do not react instantly to changes in for example taxes, wages and business prospects but take their time to make decisions and adjust behaviour. Moreover, because of information and processing lags, time often goes by before changes in economic conditions are fully recognized. There are also institutional arrangements,social and legal agreements and norms that hinder continuous adjustments of economic variables. Annual (or even biannual) wage bargaining rounds is one important example. The manufacturing of goods is usually not instantaneous but takes time, often several years in the case of projects with huge capital investments. Dynamic behaviour is also induced by the fact that many economic decisions are heavily influenced by what firms, households and the government anticipate about the future. Often expectation formation will attribute a large weight to past developments, since anticipations usually build on past experience. In macroeconomic models, dynamic behaviour often takes the form of relationships between flow and stock variables. Examples of flow variables are GDP (and the expenditure components), exports and imports, hours worked and inflation. Flow variables are measured in for example million kroner, or thousand hours worked, per year (or quarter, or month). Inflation is measured as a rate, or percentage per time period and is another example of a flow variable. Values of stock variables, in contrast, refer to particular point in time. For example, statistics of national debt often refer to the end of the year (thought it can also be given as an annual average of end-of-quarter or end-of-month observations). Prices indices are also examples of stock variables. They represent the cost of buying a basket of goods with reference to a particular time period. The annual consumption price index is a stock variable which is obtained as the average of the 12 monthly indices (each being a stock variable). Billion kroner Norwegian net foreign debt Figure 1.1: The Norwegian current account (upper panel) and net foreign deb (lower panel)t. Quarterly data 1980(1)-2003(4) As noted, economic dynamics often arise from the combination of flow and stock variables, and in order to substantiate that a little more we consider a few examples. First consider a nation s net foreign debt. In principle, the dynamic behaviour of debt (a stock) is linked to the value of the current account (flow) in the following way debt = current account +lastperiodsdebt +corrections. Hence if the current account is zero over the period, there are no corrections of the debt value (typically due to financial transactions), this periods net foreign debt will be equal to last period s debt. However, and ignoring corrections for simplicity, if there is a primary account surplus for some time, this will lead to a gradual reduction of debt or an increase in the nation s net wealth. Conversely, a consistent current account deficit raises a nation s debt. Figure 1.1 shows the development of the Norwegian current account and of Norwegian foreign deb. At the start of the period, Norway s net debt (a stock variable) was hovering at around 100 billion, despite a current account surplus (albeit small) in the early 1980s. Evidently, there was a substantial debt stemming from the 1970s which the nation s net financial savings (a flow) had not yet been able to wipe out. In most of the quarters in the years , the current account deficit is changed to a deficit, and, as one would expect from the debt equation above, debt is seen to increase again until it peached in the first quarters of Later in the 1990s, the current account surplus returned, and therefore the net debt was gradually reduced. 1

5 1.1. INTRODUCTION 3 4 CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS Later, at the end of the millennium, surpluses grew to unprecedented magnitudes, up to 75 billion per quarter, resulting in a sharp build up of net financial wealth, accumulating to 750 billion kroner at the end of the period. The theory of economic growth provides another example of the importance of stock and flow dynamics. For example, the level of production (a flow variable) in the economy depends on the size of the labour stock (literary speaking), and the capital stock. Due to the phenomenon of capital depreciation, some of today s production needs to be saved just in order to keep the capital stock intact in the first period of the future. Moreover, due to population growth (and a declining marginal product of labour), next period s capital stock will have to be larger than it is today if we want to avoid that output per capita declines in the next period. Hence, economic growth in terms of GDP per head necessiates that the flow of net investments (gross investments minus capital depreciation) is positive. In line with this, the dynamic equation of the capital stock is written as C 2 C 1 INC2 S 1 W 1 C 3 K t = K t 1 + J t D t where K t denotes the capital stock at the end of period t, J t is the flow of gross investments during period t, andd t is capital depreciation (also a flow). A much used assumption is that D t is proportional to the pre-existing capital stock, i.e., D t = δk t 1 where the rate of capital depreciation δ is positive number which is less than one. This gives a well known expression for the development of the capital stock K t =(1 δ)k t 1 + J t, (1.1) which plays an important role, not only in growth theory, but in real business cycle theory and intertemporal macroeconomics in general. A third example is the dynamic process between private consumption, savings and wealth, which is illustrated in Figure 1.2. C 1, near the centre of the picture represents private consumption today. Private savings today, S 1 in the picture, is defined as current income minus consumption, hence today s consumption affects S 1 as indicated by the line between C 1 and S 1. Suppose, despite households best attempt to balance income and consumption, today s savings S 1 turnedouttobetoo low. Rational consumers will then seek to correct this in the next period, by cutting back on consumption. This is an example of a dynamic relationship between two flowvariables: today ssavingiscausingtomorrow sconsumption,asindicatedby the line between S 1 and C 2, representing tomorrow s consumption. But the dynamic effect of today s saving doesn t necessarily end there. Household (financial) wealth is a stock variable which is given by last periods wealth plus current savings, as illustrated by the line from S 1 to W 1. Experience tells us that wealth may have separate effect on future consumption, and this possibility is captured by the line from W 1 to C 2 in the picture. This again illustrates dynamics between flow (S 1 and C 2 )andstock(w 1 )variables. Of course, tomorrow s eventual private consumption depends of a host of other factors, for example period 2 income which is denoted INC 2 in the picture. Moreover, C 1 S 2 Figure 1.2: A dynamic process involving consumption, savings and wealth. and INC 2 together determine tomorrow s savings S 2. By now the causal chain in closed: S 2 is going to affect C 3 either directly or through W 2 and so on. Yet another case of stock-flow dynamics is the relationship between wages and unemployment, which we will discuss in detail in chapter 2. The rate of unemployment is a stock variable which influences wage growth (a flow variable). At the same time, the rate of unemployment depends on accumulated wage growth which determines the real wage level. Similar linkages exist between nominal and real exchange rates, and provide one of the key dynamic mechanisms in the macroeconomic models of the national economy that we encounter later in the course. Because dynamics are a fundamental feature of the macroeconomy, all serious policy analysis is based on a dynamic approach. Hence, the officials responsible for fiscal and monetary policy use dynamic models as an aid in their decision process. In recent years, monetary policy had taken a more prominent and important role in activity regulation, and as we will explain later in the course, central banks in many countries have defined the rate of inflation as the target variable of economic policy. The instrument of monetary policy nowadays is the central banks sight deposit rate, i.e., the interest rate on banks deposits in the central bank. However, no central bank hopes for an immediate and strong effect on the rate of inflation after a change in the interest rate. Rather, because of the many dynamic effects triggered by a change in the interest rate, central bank governors prepare themselves to wait a substantial amount of time before the full effect of the interest rate change hits the target variable. The following statement from the web pages of Norges Bank [The Norwegian Central Bank] is typical of many central banks view:

6 1.2. STATIC AND DYNAMIC MODELS, AN EXAMPLE 5 A substantial share of the effects on inflation of an interest rate change will occur within two years. Two years is therefore a reasonable time horizon for achieving the inflation target of per cent1 One important goal of this course is to learn enough about dynamic modeling to be able to understand the economic meaning of like this and similar statements, and to be able to form an opinion about their realism. The quotation from Norges Bank is interesting because it demonstrates that serious policy decisions are based on what beliefs the governing bodies have about the response lag between a policy change and the effect on the target variable. Formalization of such beliefs require that we develop some conceptual tools. A first step in that direction is taken in section 1.2 and 1.3, where we use a consumption function example to introduce an important class of dynamic models called autoregressive distributed lag models, and also the concept of the dynamic multiplier. The dynamic multiplier is a key concept in this course, and once you get a good grip on it, you also have a powerful tool which allows you to calculate the dynamic effects of policy changes (and of other exogenous shocks for that matter) on important variables like consumption, unemployment, inflation or other variables of your interest. Section 1.4 then established the general notation. Section 1.5 introduces a typology of dynamic equations which are useful in economics, while section 1.6 extends the basic framework in directions that are important in applications, and brieflypresents several substantive examples. After having emphasized the difference between static and dynamic models at the start of the chapter, section 1.7 and 1.8 show that there is after all a way of reconciling the two approaches (section 1.7), and that for some purposes we can be comfortable with using a static model formulation as long as we are aware of its limitations (section 1.8). Section 1.9 shows briefly that underlying both dynamic policy analysis, and the correspondence between dynamic and static formulations, is the nature of the solution of dynamic models. Finally, section 1.10 sketches how the analysis can be extended to systems of equations with a dynamic specification. Finally, and looking ahead, in Chapter 2 the analytical framework of this chapter is applied to wage and price dynamics. 1.2 Static and dynamic models, an example As students of economics you will be familiar with model analysis, both graphical and algebraic. Presumably, most of the models you have used have been static, and time has not played an essential role in the model formulation or in the analysis. We therefore start by contrasting static models with models that have a dynamic formulation. 1 Similar statements can be found on the web pages of the central banks in e.g., Autralia, New- Zealand, The United Kingdom and Sweden. 6 CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS The intention with dynamic models is to describe the behaviour of economic variables over time. A variable y t is called a time series if we observe it over a sequence of time periods represented by the subscript t., for example {y T,y T 1,...,y 1 } represents the case where we have observations from period 1 to T. Usually, the simpler notation y t,t=1,..., T is used. The interpretation of the time subscript varies from case to case, it can represent a year, a quarter of a year, or a month. In macroeconomics, other time periods are also considered, such as 5-year or 10-year averages of historical data, and daily or even hourly data at the other extreme (e.g., exchange rates, stock prices, money market interest rates). We have already seen examples of the behaviour of actual time series in Figure 1.1, showing quarterly observations of the current account and of net debt in Norway. In this section we discuss in some detail an example where y t is (the logarithm) of private consumption, and we consider in detail a dynamic model of consumption. When we consider economic models to be used in an analysis of real world macro data, care must be taken to distinguish between static and dynamic models. The textbook consumption function, i.e., the relationship between real private consumption expenditure (C) and real households disposable income (INC) is an example of a static equation C t = f(inc t ), f 0 > 0. (1.2) Consumption in any period t is strictly increasing in income, hence the positive signed first order derivative f 0 which is called the marginal propensity to consume. To be able to apply the theory to observations of the real economy we also have to specify the function f(inc t ). Two of the most used functional forms in macroeconomics, are the linear and log-linear specifications. For the case of the static consumption function in (1.2), these two specifications are C t = β 0 + β 1 INC t + e t, (linear) (1.3) ln C t = β 0 + β 1 ln INC t + e t, (log-linear) (1.4) For simplicity we use the same symbols for the coefficients in the two equations. However,itisimportanttonotethatsincethevariablesaremeasuredondifferent scales million kroner at fixed prices in (1.3), the natural logarithm of fixed million kroner in (1.4) the slope coefficient β 1 has a different economic interpretation in the two models. 2 Thus, in equation (1.3), β 1 is the marginal propensity to consume and is in units of million kroner. Mathematically, β 1 in (1.3) is the derivative of real private consumption, C t with respect to real income, INC t : dc t = β dinc 1,from(1.3). t 2 In practice, this means that if LC t is privat consumption expenditure in million (or billion) kroner in period t, C t is defined as LC t/p C t where PC t is the price deflator in period t. If, for example, C t is in million 2000 kroner, this means that the base year of the consumer price index PC t is 2000 (with annual data, PC 2000 =1, with quarterly data, the annual average is 1 in 2000). Y t is defined accordingly.

7 1.2. STATIC AND DYNAMIC MODELS, AN EXAMPLE 7 In the log linear model (1.4), since both real income and real consumption are transformed by applying the natural logarithm to each variable, it is common to say that each variable have been log-transformed. By taking the differential of the log-linear consumption function we obtain (see appendix A for a short reference on logarithms): dc t dinc = β C 1 or t INC t dc t INC t = β dinc t C 1. t Hence, in equation (1.4), β 1 is interpreted as the elasticity of consumption with respect to income: β 1 represents the (approximate) percentage increase in real consumption due to a 1% increase in income. Note that the log-linear specification (1.4) implies that the marginal propensity to consume is itself a function of income (see exercise 1). In that sense, the log-linear model is the more flexible of the two functional forms and this is part of the reason for its popularity. To gain familiarity with the log-linear specification, we choose that functional form in the rest of this section, and in the next, but later we will also use the linear functional form, when that choice makes for easier exposition. We include a stochastic term e t in the two specified consumption functions. You can think of e t as a completely random variable, with mean value of zero, but which cannot be predicted from C t of Y t (or from the history of these two variables). In a way, the inclusion of e t in the models is a concession to reality, since economic theory (f(inc t ) in this case) cannot hope to capture all the vagaries of C t,which instead is represented by the stochastic term e t. Put differently, even if our theory is correct, it is only true on average. Hence, even if we know β 0, β 1 and INC t with certainty, the predicted consumption expenditure in period t will only be equal to C t on average, due to the random disturbance term e t. 3 There is another reason for including a disturbance term in the consumption function, which has to do with how we confront our theory with the time series evidence. So: let us consider real data corresponding to C t and Y t, and assume that we have a good way of quantifying the intercept β 0 and the elasticity of consumption with respect to income, β 1. You will learn about so called least-squares estimation in courses in econometrics, but intuitively, least-squares estimation is a way of finding the numbers for β 0 and β 1 that give the on average best prediction of C t for a given value of Y t. Using quarterly data for Norway, for the period 1967(1)-2002(4) the number in brackets denotes the quarter we obtain by using the least squares method: ln Ĉt = ln INC t (1.5) 3 Strictly speaking, we should use separate symbols not only for the coefficients, but also for the two disturbances. Logically, the same random variable cannot act as a disturbance in the two different functional forms. However, to economize notation, we have chosen to dodge this formality. 8 CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS ln C t ln INC t Figure 1.3: The estimated model in (1.5), see text for explanation. where the hat in Ĉt is used to symbolize the fitted value of consumption given the income level INC t. Next, use (1.4) and (1.5) to define the residual ê t : ê t =lnc t ln Ĉt, (1.6) which is the empirical counterpart of e t. In Figure 1.3 we show a cross-plot of the 140 observations of consumption and income (in logarithmic scale), with each observation marked by a +. The upward sloping line represents the linear function in equation (1.5), and for each observation we have also drawn the distance up (or down) to the line. These projections are the graphical representation of the residuals ê t. Clearly, if we are right in our arguments about how pervasive dynamic behaviour is in economics, equation (1.4) is a very restrictive formulation. For example, according to (1.4), the whole adjustment to a change in income is completed within a single period, and if income suddenly changes next period, consumer s expenditure changes suddenly too. As noted above, immediate and complete adjustment to changes in economic conditions is seldom seen, and dynamic models is designed to account for the typical lags in adjustment. A dynamic model of private consumption allows for the possibility that period t 1 income affects consumption, and that for example habit formation induces a positive relationship between period t 1 and period t consumption: ln C t = β 0 + β 1 ln INC t + β 2 ln INC t 1 + α ln C t 1 + ε t (1.7)

8 1.2. STATIC AND DYNAMIC MODELS, AN EXAMPLE 9 10 CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS Residuals of (1.7) Residuals of (1.4) stated, this is a typical finding with macroeconomic data. Judging from the estimated coefficients in (1.8), one main reason for the improved fit of the dynamic model is the lag of consumption itself, i.e., the autoregressive aspect of the equation. That the lagged value of the endogenous variable is an important explanatory variable is also a typical finding, and just goes to show that dynamic models represent essential tools for empirical macroeconomics. The rather low values of the income elasticities (0.130 and 0.08) may reflect that households find that a single quarterly change in income is too little to build on in their expenditure decisions. As we will see in the next section the results in (1.8) imply a much higher impact of a permanent change in income than of a temporary rise Dynamic multipliers Figure 1.4: Residuals of the two estimated consumptions functions (1.5), and (1.8). Theliteraturereferstothistypeofmodelasan autoregressive distributed lag model, ADL model for short. Autoregressive refers to the inclusion of ln C t 1 on the right hand side of the equation. Distributed lag refers to the presence of current and lagged income in the model. Arguably, a more precise formulation would be presence of current and/or lagged income in the model, since even with β 1 =0 or β 2 =0we would still refer to the equation as a an ADL model. Only when β 1 = β 2 =0does the model reduce to a purer autoregressive model (AR for short). With reference to Figure 1.2 above, note that equation (1.7) is compatible with the flow dynamics between consumption and income. In order to accommodate the effects that operate via wealth accumulation, we would of course need a more complicated model. How can we evaluate the claim that the ADL model in equation (1.7) gives a better description of the data than the static model? A full answer to that question would take us into the realm of econometrics, but intuitively, one indication would be if the empirical counterpart to the disturbance of (1.7) are smaller and less systematic than the errors of equation (1.4). To test this, we obtain the residual ˆε t, again using the method of least squares to find the best fit ofln C t according to the dynamic model: ln Ĉt = ln INC t ln INC t ln C t 1 (1.8) Figure 1.4 shows the two residual series ê t (static model) and ˆε t (ADL model), and it is immediately clear that the dynamic model in (1.8) is a much better description of the behaviour of private consumption than the static model (1.5). As already The quotation from Norges Bank s web pages on monetary policy shows that the Central Bank has formulated a view about the dynamic effects of a change in the interest rate on inflation. In the quotation, the Central Bank states that the effect will take place within two years, i.e., 8 quarters in a quarterly model of the relationship between the rate of inflation and the rate of interest. That statement may be taken to mean that the effect is building up gradually over 8 quarters and then dies away quite quickly, but other interpretations are also possible. In order to inform the public more fully about its view on the monetary policy transmission mechanism (see topic 5 in our course), the Bank would have to give a more detailed picture of the dynamic effects of a change in the interest rate. Similar issues arise whenever it is of interest to study how fast and how strongly an exogenous perturbation or a policy change affect the economy, for example how private consumption is likely to be affected by a certain amount of tax-cut. Thekeyconceptthatweneedtointroduce in detail is called the dynamic multiplier. In order to explain dynamic multipliers, we first show what our estimated consumption function implies about the dynamic effect of a change in income, see section We then derive dynamic multipliers using a general notation for autoregressive distributed lag models, in section Dynamic effects of increased income on consumption We want to consider what the estimated model in (1.8) implies about the dynamic relationship between income and consumption. For this purpose there is no point to distinguish between fitted and actual values of consumption, so we drop the ˆ above C t. Assume that income rises by 1% in period t, soinsteadofinc t we have INCt 0 = INC t ( ). Since income increases, consumption also has to rise. Using (1.8) we have ln(c t (1 + δ c,0 )) = ln(inc t ( )) ln INC t ln C t 1

9 1.3. DYNAMIC MULTIPLIERS 11 where δ c,0 denotes the relative increase in consumption in period t, thefirst period of the income increase. Using the approximation ln(1+δ c,0 ) δ c,0 when 1 <δ c,0 < 1, and noting that ln C t ln INC t 0.08 ln INC t ln C t 1 =0, we obtain δ c,0 = as the relative increase in C t. 4 In other words, the immediate effect of a one percent increase in INC is a 0.13% rise in consumption. The effect on consumption in the second period depends on whether the rise in income is permanent or only temporary. It is convenient to first consider the dynamic effects of a permanent shock to income. Note first that equation (1.8) holds also for period t +1,hencewehave before the shock, and ln(c t+1 )= ln INC t ln INC t ln C t ln(c t+1 (1 + δ c,1 )) = ln(inc t+1 ( )) +0.08(ln INC t ( )) ln(c t (1 + δ c,0 )), after the shock. Remember that in period t +1,notonlyINC t+1 has changed, but also INC t and period t consumption (by δ c,0 ). Using (again) the approximation that ln(1 + δ c,i ) δ c,i (i =0, 1), the second multiplier is found to be: δ c,1 = {ln(c t+1 ) ln INC t ln INC t 0.79 ln C t } = , or 0.3% (remember that the expression inside the brackets is zero!). By the same way of reasoning, we find that the percentage increase in consumption in period t+2 is 0.46% (formally δ c,2 100). Since δ c,0 measures the direct effect of a change in INC, it is usually called the impact multiplier, and can be defined directly by taking the partial derivative ln C t / ln INC t in equation (1.8) (more on the relationship between derivative and multipliers in section 1.4 below). The dynamic multipliers δ c,1,δ c,2,...δ c, are in their turn linked by exactly the same dynamics as in equation (1.8), namely δ c,j =0.13δ inc,j +0.08δ inc,j δ c,j 1,forj =1, 2,.... (1.9) 4 If you are unfamiliar with the the approximation ln(1 + δ c,0) δ c,0, seeappendixaandthe references given there. Note also that the if we had instead used a linear ADL model (with no log transforms of C t or INC t) in this example, there would have been no need for the approxiation, and the multipliers would have been exact. Finally, you may note that in practice the issue of the approximation does not arise, since the multiplier will be obtained by computer simulation which will give the exact numbers in an effcient way, see e.g., section below. 12 CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS For example, for j = 3, and setting δ inc,3 = δ inc,2 = 0.01 since we consider a permanent rise in income, we obtain δ c,3 = = or 0.57% in percentage terms. In this example, the multipliers increase from period to period, but the positive increment (δ c,j δ c,j 1 ) is becoming smaller and smaller. From equation (1.9) this is seen to be due to the multiplication of the multiplier δ c,j 1 by the coefficient of the autoregressive term, which is less than 1. Eventually, the sequence of multipliers converges to what we refer to as the long-run multiplier. Hence, in (1.9) if we set δ c,j = δ c,j 1 = δ c,long run we obtain δ c,long run = =0.01, meaning that according to the estimated model in (1.8), a 1% permanent increase in income has a 1% long-run effect on consumption. Remember that the set of multipliers we have considered so far represent the dynamic effects of a permanent rise in income. They are shown for convenience in the first column of table 1.1. If we instead consider a temporary rise in income (by 0.01), equation (1.8) implies a different sequence of multipliers: The impact multiplier is again , but the second multiplier becomes =0.0018, and the third is found to be = , so these multipliers are rapidly approaching zero, which becomes the long-run multiplier in this case. Table 1.1: Dynamic multipliers of the estimated consumption function in (1.8), percentage change in consumption after a 1 percent rise in income. Permanent 1% change Temporary 1% change Impact period period after shock period after shock period after shock long-run multiplier Note that the second multiplier of the permanent change is equal to the sum of the two first multipliers of the transitory shock ( = 0.31). This is due an an underlying algebraic relationship between the two types of multipliers: multiplier j of a permanent shock is the cumulated sum of the j first multipliers of a temporary shock. Specifically, if we supplement the multipliers in the column to the right with a few more periods and then sum the whole sequence, you will find that the sum is close to 1, which is the long-run multiplier of the permanent change. A relationship like this always holds, no matter what the long-run effect of the permanent change may be. Heuristically, this is because the effect of a permanent change in an explanatory

10 1.3. DYNAMIC MULTIPLIERS CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS 1.4 General notation of the ADL model Percentage change Temporary change in income Percentage change Permanent change in income As noted in the consumption function example, the impact multiplier is identical to the partial derivative of C t with respect to INC t. We now establish more formally that also the second, third and higher order multipliers can be interpreted as derivatives. At this stage it is convenient to introduce a more general notation for the autoregressive distributed lag model. In (1.10), y t is the endogenous variable while x t and x t 1 make up the distributed lag part of the model: Percentage change Period Dynamic consumtion multipliers (temporary change in income) Period Percentage change Period Dynamic consumption multipliers (permanent change in income) Period Figure 1.5: Temporary and permanent 1 percent changes in income with associated dynamic multipliers of the consumption function in (1.8). variable can be viewed as the sum of the changes triggered by a temporary change. In this sense, the dynamic multipliers of a temporary change are the more fundamental of the two: If we first calculate the effects of a temporary shock, the dynamic effects of a permanent shock can be calculated afterwards by cumulative summation. For this reason, some authors reserve the term dynamic multiplier for the effects of a temporary change, and they use a different term cumulated multipliers for the dynamic effects of a permanent change. However, as long as we are clear about which kind of shock we have in mind, no misunderstandings should result from using the term dynamic multipliers in both cases. Figure 1.5 shows graphically the two classes of dynamic multipliers for our consumption function example. 5 Panel a) show the temporary change in income, and below it, in panel c), you find the graph of the consumption multipliers. Correspondingly, panel b) and d) show the graphs with permanent shift in income and the corresponding (cumulated) dynamic multipliers. 6 5 In this book we often use figures with multiple graphs, such as Figure 1.5. In the text, we refer to the panels as a), b) etc, according to the rule a c for a 2 2 figure. In the case of a 3 3 figure for example, c) denotes the third panel of the first row, while e) is the third panel of the second row. 6 These graphs were constructed using PcGive and GiveWin, but it is of course possible to use b d y t = β 0 + β 1 x t + β 2 x t 1 + αy t 1 + ε t. (1.10) Inthesamewayasbefore,ε t symbolizes the small and random part of y t which is unexplained by x t, x t 1 and y t 1. In many applications, as in the consumption function example, y and x are in logarithmic scale, due to the specification of a log-linear functional form. However, in other applications, different units of measurement are the natural ones to use. Thus, depending of which variables we are modelling, y and x canbemeasuredinmillion kroner, or in thousand persons or in percentage points. Mixtures of measurement are also often used in practice: for example in studies of labour demand, y t may denote the number of hours worked in the economy (or by an individual) while x t denotes real wage costs per hour. The measurement scale does not affect the derivation of the multipliers, but care must be taken when interpreting and presenting the results. Specifically, only when both y and x are in logs, are the multipliers directly interpretable as percentage changes in y following a 1% increase in x, i.e., they are (dynamic) elasticities. To establish the connection between dynamic multipliers and the derivatives of y t, y t+1, y t+2,..., it is convenient to think of x t, x t+1, x t+2,,... as functions of a continuous variable h. Whenh changes permanently, starting in period t,we have x t / h > 0, while there is no change in x t 1 and y t 1 since those two variables are predetermined in the period when the shock kicks in. Since x t is a function of h, so is y t,andtheeffect of y t of the change in h is found as y t h = β x t 1 h. In the outset x t / h can be any number, and in the consumption function example we used 0.01 to denote a small change income, but in general it is convenient to evaluate the multipliers for the case of x t / h =1(it is customary to refer to this as a unit change ). Hence the first multiplier is simply y t h = β 1. (1.11) The second multiplier is found by considering the equation for period t +1, i.e., Excel or other programs. y t+1 = β 0 + β 1 x t+1 + β 2 x t + αy t + ε t+1.

11 1.4. GENERAL NOTATION OF THE ADL MODEL 15 and calculating the derivative y t+1 / h. Note that, due to the change in h occurring already in period t, bothx t+1 and x t have changed, i.e., x t+1 / h > 0 and x t / h > 0. Finally, we need to keep in mind that also y t is a function of h, hence: y t+1 h = β x t+1 1 h + β x t 2 h + α y t (1.12) h Again, considering a unit-change, x t / h = x t+1 / h =1, and using (1.11), the second multiplier can be written as y t+1 h = β 1 + β 2 + αβ 1 = β 1 (1 + α)+β 2 (1.13) To find the third derivative, consider y t+2 = β 0 + β 1 x t+2 + β 2 x t+1 + αy t+1 + ε t+2. Using the same logic as above, we obtain y t+2 h x t+2 = β 1 h + β x t+1 2 h + α y t+1 h = β 1 + β 2 + α y t+1 h = β 1 (1 + α + α 2 )+β 2 (1 + α) (1.14) where the conventional unit-change, x t / h = x t+1 / h =1, is used in the second line, and the third line is the result of substituting y t+1 / h by the right hand side of (1.13). Comparison of equation (1.12) with the first line of (1.14) shows that there is a clear pattern: The third and second multipliers are linked by exactly the same form of dynamics that the govern the y variable itself. This also holds for higher order multipliers, and means that the multipliers can be computed recursively: Once we have found the second multiplier, the third can be found easily by using the second line of (1.14). Table 1.2 contains a summary of the results we have obtained. In the table, we use the notation δ j (j =0, 1, 2,...) for the multipliers. For, example δ 0 is identical to y t / h, andδ 2 is identical to the third multiplier, y t+2 / h above. In general, because the multipliers are linked recursively, multiplier number j +1is given as δ j = β 1 + β 2 + αδ j 1,forj =1, 2, 3,... (1.15) In the consumption function example, we saw that as long as the autoregressive parameter is less than one, the sequence of multipliers is converging towards a long run multiplier. In this more general case, the condition needed for the existence of a long-run multiplier is that α is less than one in absolute value, formally 1 <α<1. In later sections, namely 1.7 and 1.9 specifically, this condition is explained in more detail. For the present purpose, we simply assume that the condition holds, and 16 CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS Table 1.2: Dynamic multipliers of the general autoregressive distributed lag model. ADL model: y t = β 0 + β 1 x t + β 2 x t 1 + αy t 1 + ε t. Permanent unit change in x (1) Temporary unit change in x (2) 1. multiplier: δ 0 = β 1 δ 0 = β 1 2. multiplier: δ 1 = β 1 + β 2 + αδ 0 δ 1 = β 2 + αδ 0 3. multiplier: δ 2 = β 1 + β 2 + αδ 1 δ 2 = αδ 1... j+1 multiplier δ j = β 1 + β 2 + αδ j 1 δ j = αδ j 1 long-run (3) δ long run = β 1 +β 2 1 α 0 notes: (1) As explained in the text, x t+j / h =1, j =0, 1, 2,... (2) x t / h =1, x t+j / h =0, j =1, 2, 3,... (3) Subject to the restriction 1 <α<1. define the long-run multiplier as δ j = δ j 1 = δ long run. Using (1.15), the expression for δ long run is found to be δ long run = β 1 + β 2,if 1 <α<1. (1.16) 1 α Clearly, if α =1, the expression does not make sense mathematically, since the denominator is zero. Economically, it doesn t make sense either since the long run effect of a permanent unit change in x is an infinitely large increase in y (if β 1 + β 2 > 0). Thecaseofα = 1, mayatfirst sight seem to be acceptable since the denominator is 2, not zero. However, as explained in section 1.9 below, the dynamics is unstable also in this case, meaning that the long run multiplier is not well definedforthecaseofα = 1. Hence, while we can use the table to calculate dynamic multipliers also for the cases where the absolute value of α is equal to or larger than unity, the long-run multiplier of a permanent change in x does not exist in this case. Correspondingly: the multipliers of a temporary change do not converge to zero in the case where 1 <α<1 does not hold. Notice that, unlike α =1, thecaseofα =0is unproblematic. This restriction, which excludes the autoregressive term y t 1 from the model, only serves to simplify the multiplier analysis. All of the dynamic response in of y then due to the distributed lag in the explanatory variable, anditisreferredtobyeconomistsasthe distributed lag model, denoted DL-model.

12 1.5. A TYPOLOGY OF LINEAR MODELS 17 Table 1.3: A model typology. Type Equation Restrictions ADL y t = β 0 + β 1 x t + β 2 x t 1 + αy t 1 + ε t. None Static y t = β 0 + β 1 x t + ε t. β 2 = α =0 DL y t = β 0 + β 1 x t + β 2 x t 1 + ε t α =0 Differenced data 1 y t = β 0 + β 1 x t + ε t β 2 = β 1, α =1 ECM y t = β 0 + β 1 x t +(β 1 + β 2 )x t 1 None +(α 1)y t 1 + ε t Homogenous ECM y t = β 0 + β 1 x t β 1 + β 2 = (α 1) +(α 1)(y t 1 x t 1 )+ε t 1 is the difference operator, defined as z t z t z t A typology of linear models The discussion at the end of the last section suggests that if the coefficient α in the ADL model is restricted to for example 1 or to 0, quite different dynamic behaviour of y t is implied. In fact the resulting models are special cases of the unrestricted ADL model. For reference, this section gives a typology of models that are encompassed by the ADL model. Some of these model we have already mentioned, while others will appear later in the book. 7 Table 1.3 contains three columns, for model Type, defining Equation and Restrictions, where we give the coefficient restrictions that must be true for each models to be a valid simplifications of the ADL. Hence for example, for the Static model to be valid, both β 2 =0and α =0must be true (in practice this amounts to estimation of the ADL and then testing the joint hypothesis with a F-test). The danger of using a static model when the restrictions do not hold is that we get a misleading impression of the adjustment lags (the dynamic multipliers). Specifically, the response of y t to a change in x t is represented as immediate when it is in fact distributed over several periods. For the DL model to be a valid simplification of the ADL, only one coefficient restriction needs to be true, namely α =0. That said, the DL model is also a quite restrictive model of the dynamic response to a shock, since the whole adjustment of 7 In fact our typology is partial, and covers only 5 of the 9 models in the full typology of Hendry (1995). 18 CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS y t to a change in x t is completed in the course of only two periods. The fourth model in Table 1.3, called the Differenced data model is included since it is popular in modern macroeconomics. A technical motivation for the model in applied work is that, by taking the difference of y t and x t prior to estimation, the econometric problem of residual autocorrelation is reduced. However, unless the two restrictions that define the model type are empirically acceptable, choosing the Differenced data model may create more problems than it solves. For example, the regression coefficient of x t will not be a correct estimate of the impact multiplier β 1, nor of the long-run multiplier δ long run. The economic interpretation of the Differenced data model ( growth rate model if y t and x t are in logs) is also problematic, since there appears to be no cost of being out of equilibrium in terms of the levels variables. The two last models do not have the shortcomings of the Static, DL or Differenced data models. Hence, the error-correction model (ECM) is consistent with a long run relationship between y and x, and it also describes the behaviour of y t outside equilibrium. As explained in section 1.7 below, the ECM is actually a reformulation of the ADL, hence there are no restrictions involved. The Homogenous ECM has the same advantages as the ECM in terms of interpretation, but the long-run multiplier is restricted to unity. One example, which is relevant for our course, is when y t is the log of the domestic price level, and x t is the log of an index of foreign prices, denoted in domestic currency. If the hypothesis of purchasing power parity applies, that would motivate the Homogenous ECM with β = Extensions and examples In this subsection we briefly point to some important extensions of the ADL model. Second, to help solidify the understanding of the ADL framework, we provide additional economic examples Extensions The ADL model in equation (1.10) is general enough to serve as an introduction to most aspects of dynamic analysis in economics. However, the ADL model, and the dynamic multiplier analysis, can be extended in several directions to provide additional flexibility in applications. The most important extensions are: 1. Several explanatory variables 2. Longer lags 3. Systems of ADL equations In economics, more than one explanatory variable is usually needed to provide a satisfactory explanation of the behaviour of a variable y t. The ADL model (1.10) can be generalized to any number of explanatory variables. However, nothing is lost

13 1.6. EXTENSIONS AND EXAMPLES 19 in terms of understanding by only considering the case of two exogenous variables, x 1,t and x 2,t. The extension of equation (1.10) to this case is y t = β 0 + β 11 x 1,t + β 21 x 1,t 1 + (1.17) β 21 x 2,t + β 22 x 2,t 1 + αy t 1 + ε t, where β ik is the coefficient of the i 0 th lag of the explanatory variable k. The dynamic multipliers of y t can be with respect to either x 1 of x 2, the derivation being exactly the same as above. Formally, we can think of each set of multipliers as corresponding to partial derivatives. In applications, the dynamic multipliers of the different explanatory variables are often found to be markedly different. For example, if y t is the (log of) the hourly wage, while x 1 and x 2 are the rate of unemployment and productivity respectively, dynamic multipliers with respect to unemployment is usually much smaller in magnitude than the multipliers with respect to productivity, see section 2 below. Longer lags in either the x-es or in the autoregressive part of the model makes for more flexible dynamics. Even though the dynamics of such models can be quite complicated, the dynamic multipliers always exist, and can be computed by following the logic of section above. However, as such calculation quickly become tiring, practitioners use software when they work with such models (good econometric software packages includes options for dynamic multiplier analysis). In this course, we do not consider the formal analysis of higher order autoregressive dynamics. Often, an ADL equation is joined up with other equations to form dynamic system of equations. For example, in (1.17) while x 2 is truly exogenous, there is separate equation for x 1,t with y t on the right hand side. Hence, x 1,t and y t are determined in a simultaneous equations system. Another, very common case, is that the equation for x 1,t contains the lag y t 1, not the contemporaneous y t.alsointhis case is x 2,t and y t jointly determined, but in what is referred to as a recursive system of equations. In either case the true multipliers of y t with respect to the exogenous variable x 2,t, cannot be derived from equation (1.17) alone, because that would make us miss the feed-back that a change in x 2,t has on y t via the endogenous variable x 1,t. To obtain the correct dynamic multipliers of y with respect to x 1 we must use the two equation system. In this course we will not pursue the formal analysis of dynamic system is general. However, specific economicexamplesofdynamicsystem and how they are analyzed will crop up several times as we proceed. For example, section 1.10 takes the dynamic version of the Keynesian multiplier models as one specific example. Other applications of dynamic systems are found in the sections on wage-price dynamics and on economic policy analysis A few more examples There are more examples of applications of the ADL model than one can mention, so this section focuses on a few of the examples that we encounter later in the course. 20 CHAPTER 1. DYNAMIC MODELS IN MACROECONOMICS The dynamic consumption function (again) This has been the main example so far, and in section 1.3 we discussed the dynamics of a log-linear specification in detail. Of course exactly the same analysis applies to a linear functional form of the consumption function, except that the multipliers will be in units of million kroner (at fixed prices) rather than percentages. In section 1.10 the linear consumption function is combined with the general budget equation to form a dynamic system. In modern econometric work on the consumption function, more variables are usually included than just income. Hence, there are other multipliers to consider than the ones with respect to INC t. The most commonly found additional explanatory variables are wealth, the real interest rate and indicators of demographic developments The Phillips curve In Chapter 2, and several times later in the course, we will consider the so called expectations augmented Phillips curve. An example of such a relationship is π t = β 0 + β 11 u t + β 12 u t 1 + β 21 π e t+1 + ε t. (1.18) π t denotes the rate of inflation, π t =ln(p t /P t 1 ),wherep t is an index of the price level of the economy. u t is the rate of unemployment or its log. Whether u t is a rate (or percentage) or the log of a rate (percentage) is another of example of a choice between different functional forms (as in the consumption function example at the beginning of this chapter). If u t is a rate (or percentage) the Phillips curve is linear, and the effect of a small change in u t on π t is independent of the initial unemployment rate. If u t =ln(the rate of unemployment), the Phillips curve is non-linear: starting from a low level, a rise in the rate of unemployment leads to a lager reduction in π t than if the initial rate of unemployment is high. The graph of the Phillips curve, with the rate of unemployment on the horizontal axis, is thus curved towards the origin (a convex function). 8 In many countries, a convex Phillips curve is known to give a better data fit than the linear function, and it is important to keep that in mind when we frequently in this book, find it convenient to interpret u t as the rate of unemployment. In equation (1.18), the distributed lag in the rate of unemployment captures several interesting economic hypotheses. For example, if the rise in inflation following a fall in unemployment is first weak but then gets stronger, we might have that both β 11 < 0 and β 12 < 0. On the other hand, some economist have argued the opposite: 8 In Norway, Phillips curves, and also the error correction models of wage formation that we encounter below, have alternatively included a reciprocal term, i.e., 1/u t,whereu t is the rate of unemployment. This functional form is even more non-linear than the log-form. For example, beyond a certain level of unemployment, there is no more inflation reduction to be hauled from further increases of unemployment.